[Benchmark] DeepSeek-V4.1-Flash-IST-DAS-2.25bpw on RTX 5070 12GB : 10~12 tok/s (Strata 0.2.11)

Aperta Adatta ai principianti
  • Qualcosa non va?
#5,434 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 4 giorni

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
1/5
Tempo stimato
Meno di un'ora
Idoneità per principianti
70/100
Tipo di issue
Documentazione
Chiarezza
Specificata chiaramente
Stato di attività
Attiva
Stack tecnologico
cpp, linux
Ambito
machine-learning, performance

Direzione di ricerca

Il report menziona tools/strata_bench.py e rimanda allo script di benchmark in PR #4877; inizia controllando il formato per i contributi ai benchmark del repository e i report di benchmark esistenti. Aggiungi questo risultato della RTX 5070 nel formato corrispondente, quindi verifica che l’hardware, le impostazioni del modello, le misurazioni e i limiti siano rappresentati accuratamente.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

i7-12700K, 64GB DDR4 (61.3 available), RTX 5070 12GB - pcie gen5, nvme SSD, Linux Ubuntu 26.04, cuda13, Strata 0.2.11
Both CPU and GPU undervolted.
mtp-bench-3.py from PR #4877

tools/strata_bench.py

Community benchmark on NVIDIA GeForce RTX 5070

Measured on 2026-11-06 with tools/strata_bench.py (suite 1, 3 runs per test, medians with ranges). Speeds are the engine's own timings; first-token times are measured by the client over HTTP. Speed only: no correctness check was run.

Hardware and software

  • GPU and VRAM: NVIDIA GeForce RTX 5070, 12 GiB VRAM
  • CPU: 12th Gen Intel(R) Core(TM) i7-12700K (12 cores, 20 threads); installed RAM: 61.3 GiB; storage: nvme2n1 WDS100T1X0E-00AFY0 nvme; nvme1n1 WD Blue SN570 500GB nvme; nvme3n1 WDS100T1X0E-00AFY0 nvme; nvme0n1 WDS200T1X0E-00AFY0 nvme
  • OS: Linux 7.0.0-38-generic; GPU driver: 615.71.10; CUDA (driver supports up to): 13.4
  • Strata: project ?, commit not a git checkout; engine ds4-0.1
  • GPU detail (after the run, near idle): power_limit_w=300.00, power_draw_w=60.72, pcie=gen5 x16, pstate=P1, temp_c=47
  • Background workloads: not recorded (GPU power limit/draw read above)

Model and configuration

  • Model and quantization: deepseek-v4.1-flash
  • Context: 131072 tokens; parallel slots: 1
  • Settings: the install's own configuration, unchanged; thinking off for the suite; temperature 0

Method

Run by python tools/strata_bench.py against the HTTP API of a Strata server (already running); model loading is not timed. One warm-up chat and a 1K prompt are discarded first. Every prompt starts with a unique tag, so the conversation cache cannot help the fresh-prompt tests.

  • short chat: the three questions of tools/ab_engine.py, up to 100 new tokens, 3 runs
  • 4K / 16K prompts: generated word lists sized by a calibration request, 24 new tokens, 3 fresh prompts each
  • agent session: system prompt and four tools, then 5 turns that each add about 1500 words of a project file and continue from the model's own earlier answers (the way agent clients resend a conversation), 64 new tokens per turn, 3 sessions; the cache numbers come from the answer's usage (cached_tokens) and timings (prompt_n)
  • concurrency (when run): N requests sent at the same moment, total tokens over the wall time of the burst, latency over all requests

Results

Configuration Actual prompt tokens Reused tokens Generated tokens Runs Prompt tok/s median and range Decode tok/s median and range TTFT seconds median and range
short chat 32 (29-38) 0 100 (100-100) 9 not measured 10.0 (9.1-10.8) 3.16 (2.52-3.49)
4K prompt, fresh 4125 (4125-4125) 0 (0-0) 24 cap 3 246.1 (236.7-246.6) not measured 16.77 (16.74-17.44)
16K prompt, fresh 16562 (16562-16562) 0 (0-0) 24 cap 3 549.4 (534.9-550.6) not measured 30.17 (30.10-30.99)
agent session, follow-up turns 8933 (8927-8956) at the last turn see below up to 64 3 see below 13.1 (9.9-14.1) 9.60 (8.09-10.25)

Ranges are the lowest and highest of the runs; Runs counts requests that returned a token.

Agent session (conversation cache)

Cache hit over the follow-up turns: 71.2 % (71.1-71.2); earlier-turn tokens read again per session: 5 (5-5) (earlier turns that the cache did not hold; the 8055 (8055-8059) new tokens of the files are read once); first turn (nothing cached) first token: 12.56 s (12.42-12.99).

Turn (run 1) Prompt tokens Reused Read Prompt tok/s First token (s) Decode tok/s
0 602 0 602 46.7 12.993 8.3
1 2252 640 1612 195.5 8.257 10.8
2 3914 2302 1612 167.5 9.644 11.5
3 5584 3972 1612 162.8 9.914 12.6
4 7255 5643 1612 165.1 9.778 13.1
5 8927 7315 1612 157.5 10.245 13.7

Per-request data: results.json. Engine log lines from the start: engine-log-excerpt.txt (when this run started the server).

Correctness and limitations

Not measured: this is a speed suite. The conversation uses the model's own replies, so token counts differ a little between runs. Machine details come from the server, nvidia-smi/rocm-smi and the OS; background load is not recorded. Total time of the suite: 568 s.

mtp-bench-3.py

~/AI/Benchmarks/mtp-bench-3.py --url http://127.0.0.1:8090
==============================================================================
MTP benchmark — subject under test
Inference engine : Strata  (version ds4-0.1)
LLM / model      : deepseek-v4.1-flash
Request dialect  : strata-chat
Model path       : /home/andrew/projects/Strata-data/models/packs/deepseek-v4.1-flash-ist-das-2.25bpw
Endpoint         : http://127.0.0.1:8090
Engine project   : https://github.com/Niko1221/Strata
Engine settings  : drafting=off, kv=fp4, context=131072, expert_slots=377, gpu_name=NVIDIA GeForce RTX 5070
note: /health: service=strata model=deepseek-v4.1-flash max_context=131072 loaded=True
note: request dialect: POST /v1/chat/completions (Strata has no llama.cpp POST /completion — that route 404s)
note: drafting method: off (spec=None mtp_max=None lookup=None)
==============================================================================
Running 9 prompts (max_tokens=192, temperature=0.0, seed=42) via strata-chat

  code_python        pred= 192 draft=   0 acc=   0 rate=n/a tok/s=10.5
  code_cpp           pred= 192 draft=   0 acc=   0 rate=n/a tok/s=11.1
  explain_concept    pred= 192 draft=   0 acc=   0 rate=n/a tok/s=11.3
  summarize          pred= 116 draft=   0 acc=   0 rate=n/a tok/s=12.5
  qa_factual         pred=  72 draft=   0 acc=   0 rate=n/a tok/s=10.8
  translation        pred= 109 draft=   0 acc=   0 rate=n/a tok/s=12.2
  creative_short     pred= 154 draft=   0 acc=   0 rate=n/a tok/s=11.7
  stepwise_math      pred= 192 draft=   0 acc=   0 rate=n/a tok/s=12.3
  long_code_review   pred= 192 draft=   0 acc=   0 rate=n/a tok/s=10.1

Benchmark results — engine: Strata ds4-0.1 | LLM: deepseek-v4.1-flash | dialect: strata-chat
Aggregate: {
  "n_requests": 9,
  "total_predicted": 1411,
  "total_draft": 0,
  "total_draft_accepted": 0,
  "aggregate_accept_rate": null,
  "wall_s_total": 174.27,
  "total_cache_n": 0,
  "server_delta": {
    "requests": 9,
    "output_tokens": 1411,
    "drafts_offered": 0,
    "drafts_accepted": 0,
    "reused": 0,
    "prompt_ms": 58236.5,
    "decode_ms": 116013.4,
    "server_accept_rate": null,
    "server_decode_tok_s": 12.2
  },
  "per_request_accept_rate": null,
  "server_accept_rate": null
}
note: per-request draft counters matches server counters (server: 0->0 accepted of 0->0 offered)
note: decode expert cache hit rate: no request records appeared in /metrics
Lingua principale
C++
Stelle
11.6k
Fork
1k
Nessuna PR unita negli ultimi 30g

Preparare l'ambiente

Questo progetto non fornisce container di sviluppo, Dockerfile né guida per i contributori, quindi l'ambiente è a tuo carico: parti dal suo README e consulta la nostra guida al primo contributo per i passaggi generali.

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di Niko1221/Strata

Tutte le issue di Niko1221/Strata

Issue simili

Altre issue su C++

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.